from_pretrained() errors in loading state_dict for FastFit

[bartschuller]

FastFit 1.2.1

I've been trying to use FastFit for Dutch using the solution in #15 but after training export_model() errors out. Or rather, the reloading using from_pretrained() that it does.
The same happens if I use model.save_pretrained() myself and then from_pretrained().

These are the errors:

RuntimeError: Error(s) in loading state_dict for FastFit:
	While copying the parameter named "all_docs.0", whose dimensions in the model are torch.Size([86, 13]) and whose dimensions in the checkpoint are torch.Size([86, 13]), an exception occurred : ('Only Tensors of floating point and complex dtype can require gradients',).
	While copying the parameter named "all_docs.1", whose dimensions in the model are torch.Size([86, 13]) and whose dimensions in the checkpoint are torch.Size([86, 13]), an exception occurred : ('Only Tensors of floating point and complex dtype can require gradients',).